On 5/10/22 2:50 AM, Jan Kara wrote:
> Sorry for delayed reply. This has fallen through the cracks...
>
> On Thu 28-04-22 13:16:19, Stefan Roesch wrote:
>> On 4/28/22 10:47 AM, Jan Kara wrote:
>>> On Tue 26-04-22 10:43:32, Stefan Roesch wrote:
>>>> This change adds support for async write throttling in the function
>>>> balance_dirty_pages(). So far if throttling was required, the code was
>>>> waiting synchronously as long as the writes were throttled. This change
>>>> introduces asynchronous throttling. Instead of waiting in the function
>>>> balance_dirty_pages(), the timeout is set in the task_struct field
>>>> bdp_pause. Once the timeout has expired, the writes are no longer
>>>> throttled.
>>>>
>>>> - Add a new parameter to the balance_dirty_pages() function
>>>> - This allows the caller to pass in the nowait flag
>>>> - When the nowait flag is specified, the code does not wait in
>>>> balance_dirty_pages(), but instead stores the wait expiration in the
>>>> new task_struct field bdp_pause.
>>>>
>>>> - The function balance_dirty_pages_ratelimited() resets the new values
>>>> in the task_struct, once the timeout has expired
>>>>
>>>> This change is required to support write throttling for the async
>>>> buffered writes. While the writes are throttled, io_uring still can make
>>>> progress with processing other requests.
>>>>
>>>> Signed-off-by: Stefan Roesch <shr@xxxxxx>
>>>
>>> Maybe I miss something but I don't think this will throttle writers enough.
>>> For three reasons:
>>>
>>> 1) The calculated throttling pauses should accumulate for the task so that
>>> if we compute that say it takes 0.1s to write 100 pages and the task writes
>>> 300 pages, the delay adds up to 0.3s properly. Otherwise the task would not
>>> be throttled as long as we expect the writeback to take.
>>>
>>> 2) We must not allow the amount of dirty pages to exceed the dirty limit.
>>> That can easily lead to page reclaim getting into trouble reclaiming pages
>>> and thus machine stalls, oom kills etc. So if we are coming close to dirty
>>> limit and we cannot sleep, we must just fail the nowait write.
>>>
>>> 3) Even with above two problems fixed I suspect results will be suboptimal
>>> because balance_dirty_pages() heuristics assume they get called reasonably
>>> often and throttle writes so if amount of dirty pages is coming close to
>>> dirty limit, they think we are overestimating writeback speed and update
>>> throttling parameters accordingly. So if io_uring code does not throttle
>>> writers often enough, I think dirty throttling parameters will be jumping
>>> wildly resulting in poor behavior.
>>>
>>> So what I'd probably suggest is that if balance_dirty_pages() is called in
>>> "async" mode, we'd give tasks a pass until dirty_freerun_ceiling(). If
>>> balance_dirty_pages() decides the task needs to wait, we store the pause
>>> and bail all the way up into the place where we can sleep (io_uring code I
>>> assume), sleep there, and then continue doing write.
>>>
>>
>> Jan, thanks for the feedback. Are you suggesting to change the following
>> check in the function balance_dirty_pages():
>>
>> /*
>> * Throttle it only when the background writeback cannot
>> * catch-up. This avoids (excessively) small writeouts
>> * when the wb limits are ramping up in case of !strictlimit.
>> *
>> * In strictlimit case make decision based on the wb counters
>> * and limits. Small writeouts when the wb limits are ramping
>> * up are the price we consciously pay for strictlimit-ing.
>> *
>> * If memcg domain is in effect, @dirty should be under
>> * both global and memcg freerun ceilings.
>> */
>> if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
>> (!mdtc ||
>> m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
>> unsigned long intv;
>> unsigned long m_intv;
>>
>> to include if we are in async mode?
>
> Actually no. This condition is the one that gives any task a free pass
> until dirty_freerun_ceiling(). So there's no need to do any modification
> for that. Sorry, I've probably formulated my suggestion in a bit confusing
> way.
>
>> There is no direct way to return that the process should sleep. Instead
>> two new fields are introduced in the proc structure. These two fields are
>> then used in io_uring to determine if the writes for a task need to be
>> throttled.
>>
>> In case the writes need to be throttled, the writes are not issued, but
>> instead inserted on a wait queue. We cannot sleep in the general io_uring
>> code path as we still want to process other requests which are affected
>> by the throttling.
>
> Probably you wanted to say "are not affected by the throttling" in the
> above.
>
Yes, that's correct.
> I know that you're using fields in task_struct to propagate the delay info.
> But IMHO that is unnecessary (although I don't care too much). Instead we
> could factor out a variant of balance_dirty_pages() that returns 'pause' to
> sleep, 0 if no sleeping needed. Normal balance_dirty_pages() would use this
> for pause calculation, places wanting async throttling would only get the
> pause to sleep. So e.g. iomap_write_iter() would then check and if returned
> pause is > 0, it would abort the loop similary as we'd abort it for any
> other reason when NOWAIT write is aborted because we need to sleep. Iouring
> code then detects short write / EAGAIN and offloads the write to the
> workqueue where normal balance_dirty_pages() can sleep as needed.
>
> This will make sure dirty limits are properly observed and we don't need
> that much special handling for it.
>
I like the idea of factoring out a function out balance_dirty_pages(), however
I see two challenges:
- the write operation has already completed at this point,
- so we can't really sleep on its completion in the io-worker in io-uring
- we don't know how long to sleep in io-uring
Currently balance_dirty_pages_ratelimited() is called at the end of the function
iomap_write_iter(). If the function balance_dirty_pages_ratelimited() would instead
be called at the beginning of the function iomap_write_iter() we could return -EAGAIN
and then complete it in the io-worker.
I'm not sure what the implications are of moving the function call to the beginning of
the function iomap_write_iter().
> Honza
